Home
last modified time | relevance | path

Searched refs:vsum0xEF (Results 1 – 24 of 24) sorted by relevance

/external/XNNPACK/src/qs8-gemm/gen/
D1x16c8-minmax-neon-mull-padal.c127 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
131 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
159 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal() local
160 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mull_padal()
D1x16c16-minmax-neon-mlal-padal.c143 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
147 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
175 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
176 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c216 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
220 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
248 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
249 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D2x16c8-minmax-neon-mull-padal.c182 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
194 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
226 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal() local
227 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mull_padal()
D2x16c16-minmax-neon-mlal-padal.c214 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
226 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
258 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
259 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c237 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
257 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
293 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal() local
294 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c321 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
333 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
365 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
366 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c285 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
305 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
341 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
342 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c292 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
320 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
360 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal() local
361 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mull_padal()
D3x16c8-minmax-neon-mlal-padal.c426 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
446 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
482 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
483 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c16-minmax-neon-mlal-padal.c356 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
384 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
424 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
425 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c531 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
559 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
599 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
600 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_gemm_minmax_ukernel_4x16c8__neon_mlal_padal()
/external/XNNPACK/src/qs8-igemm/gen/
D1x16c8-minmax-neon-mull-padal.c141 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
145 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
173 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal() local
174 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mull_padal()
D1x16c16-minmax-neon-mlal-padal.c157 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
161 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
189 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal() local
190 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_1x16c16__neon_mlal_padal()
D1x16c8-minmax-neon-mlal-padal.c230 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
234 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
262 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal() local
263 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_1x16c8__neon_mlal_padal()
D2x16c8-minmax-neon-mull-padal.c198 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
210 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
242 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal() local
243 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mull_padal()
D2x16c16-minmax-neon-mlal-padal.c230 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
242 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
274 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal() local
275 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_2x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mull-padal.c255 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
275 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
311 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal() local
312 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mull_padal()
D2x16c8-minmax-neon-mlal-padal.c337 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
349 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
381 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal() local
382 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_2x16c8__neon_mlal_padal()
D3x16c16-minmax-neon-mlal-padal.c303 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
323 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
359 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal() local
360 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_3x16c16__neon_mlal_padal()
D4x16c8-minmax-neon-mull-padal.c312 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
340 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
380 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal() local
381 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mull_padal()
D4x16c16-minmax-neon-mlal-padal.c376 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
404 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
444 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal() local
445 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_4x16c16__neon_mlal_padal()
D3x16c8-minmax-neon-mlal-padal.c444 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
464 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
500 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal() local
501 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_3x16c8__neon_mlal_padal()
D4x16c8-minmax-neon-mlal-padal.c551 const int32x4_t vsum0xEF = vpaddq_s32(vacc0x14, vacc0x15); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
579 int32x4_t vacc0xCDEF = vpaddq_s32(vsum0xCD, vsum0xEF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()
619 const int32x2_t vsum0xEF = vpadd_s32(vpsum0xE, vpsum0xF); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal() local
620 int32x4_t vacc0xCDEF = vcombine_s32(vsum0xCD, vsum0xEF ); in xnn_qs8_igemm_minmax_ukernel_4x16c8__neon_mlal_padal()